Kobin

The problem: Brands bleed identity at scale. Once multiple designers, agencies, and regional teams start shipping social posts, emails, and ads, creatives slowly drift off-brand — wrong colors, off-tone copy, flash-sale urgency for a calm brand. Catching this today is a manual review bottleneck. What TasteDrift does: TasteDrift makes brand compliance automatic and explainable. Everything revolves around one idea: the design.md — a markdown design system that serves as a reasoning artifact both humans and models can read. Brand intake from a URL. Paste a website and TasteDrift scrapes it — colors from its stylesheets, fonts, imagery, headings, and copy — then GLM 5.2 (Vultr Serverless Inference) synthesizes the brand's canonical design.md: identity, palette, typography feel, visual rules, voice, content themes, approved patterns, and explicit "avoid" rules. Creative understanding. Upload a PNG/JPG creative and NVIDIA Nemotron Omni's vision reasoning reads it — observed palette, copy, layout density, CTA style, campaign intent, likely audience — and writes the creative's own design.md. The audit. An AI rubric compares the two documents and returns a 0–100 drift score with a verdict (Approved / Needs Review / Rejected), concrete findings, actionable recommendations, and the full reasoning trace. Reviewers can also chat with TasteDrift about any brand or creative — answers are grounded in the actual design.md documents and audit results. Everything is inspectable in a Figma-style workspace: creatives cluster on a canvas by verdict, and an inspector shows both design.md documents, the reasoning, and the decision side by side. Engineering highlights: Every AI path has a deterministic fallback — regex-based brand signal extraction, OCR (Tesseract) plus pixel-math palette analysis, and a rule-based audit — so the product works end-to-end even with no API keys, and every verdict records its provenance (GLM / Nemotron / deterministic). The audit provider is switchable via a single env var. 127 unit tests plus live Playwright end-to-end tests cover the whole loop. Stack: Next.js 15 (App Router + server actions), React 19, Prisma + SQLite, Tailwind, Vultr Serverless Inference (GLM 5.2), NVIDIA API Catalog (Nemotron Omni), Vitest + Playwright. Team: Arham, Osama, Taha, Umar.